On August 27, 2026, a team led by Professor Jay A. Stein from the Massachusetts Institute of Technology (MIT) published a new paper in PNAS, introducing a machine learning framework called PottsMPNN. Developed by MIT’s Department of Biology, this framework aims to design proteins beyond natural sequence constraints. Unlike traditional methods that focus solely on replicating evolved sequences, PottsMPNN improves sequence generation capabilities and enhances the accuracy of predicting the stability impact of mutations by incorporating physical principles and evolutionary-related sequence training. This framework enables AI to “see” how multiple sequences can fold into the same structure, thereby designing entirely new proteins that are not similar to natural ones but still have feasible structures.
Professor Jay A. Stein’s team published a new paper in PNAS, proposing the PottsMPNN machine learning framework to design proteins beyond natural sequences. Developed by the Department of Biology, this framework improves sequence generation capabilities and enhances the accuracy of predicting the impact of mutations on stability by incorporating physical principles and evolutionary-related sequence training. Unlike traditional methods that focus solely on whether the model can reproduce evolutionally selected sequences, PottsMPNN aims to allow AI to “see” how multiple sequences can fold into the same structure, thereby designing entirely new proteins that are not similar to natural ones but have viable structures.